SSCST-Net：基于空谱通道选择和Transformer的高光谱图像分类模型

SSCST-Net: Hyperspectral Image Classification Model Based on Spatial-Spectral Channel Selection and Transformer

摘要: 近年来，基于Transformer的深度学习模型在高光谱图像分类任务中取得了显著的性能提高，但是由于高光谱图像具有丰富且连续的光谱特征，不可避免地存在光谱特征冗余问题，从而导致分类精度下降，同时Transformer对局部空间特征的提取存在一定的改善空间。为此，文章提出了一种基于空谱通道选择和Transformer的高光谱图像分类网络。首先，利用自适应的空谱通道选择方法来提取类别边缘特征，并保留具有高度类别特征表达能力的光谱通道，从而解决Transformer对局部空间特征关注不足和特征冗余的问题，再利用Transformer来提取空间光谱特征并进行分类。实验结果表明，该模型在Indian Pines和University of Pavia上分别取得98.45%和99.62%的整体分类精度。

Abstract: In recent years, Transformer-based deep learning models have achieved remarkable performance improvements in hyperspectral image classification tasks. However, due to the rich and continuous spectral characteristics of hyperspectral images, spectral feature redundancy inevitably exists, leading to reduced classification accuracy. Additionally, Transformer architecture still has room for improvement in extracting local spatial features. To address these issues, this paper proposes a hyperspectral image classification network based on spatial-spectral channel selection and Transformer. First, an adaptive spatial-spectral channel selection method is employed to extract class-discriminative edge features while preserving spectral channels with high class-specific representational capacity. This approach mitigates both the insufficient attention to local spatial features in the Transformer and the feature redundancy problem. Subsequently, the Transformer is utilized to capture spatial-spectral features for classification. Experimental results demonstrate that the proposed model achieves overall classification accuracies of 98.45% and 99.62% on the Indian Pines and University of Pavia datasets, respectively.